Integrated Sky-Ground Monitoring Method for Monitoring Bursaphelenchus xylophilus

Abstract

The present invention discloses an integrated sky-ground monitoring method for monitoring pine wilt disease, including: S1. First, extract diseased dead trees through UAV images, extract patches of diseased dead trees through satellite images, and conduct on-site marking, positioning and measurement of dead trees through ground surveys; S2. Obtain high-spatial-resolution satellite remote sensing data and UAV remote sensing data as data sources respectively, and combine ground survey data with forest resource second-class survey data to jointly form a data set of the monitoring area; S3. Establish samples of diseased wood through the above data, and the samples include spectral characteristics, texture characteristics and geometric characteristics of pine wilt disease dead trees; S4. According to the samples, establish the spatial distribution information and location information of pine wilt disease dead trees. The present invention has high efficiency in screening epidemic areas, can achieve large-area and high-efficiency monitoring, convenient acquisition of UAV remote sensing data, high recognition accuracy, and guaranteed accuracy verification.

Description

Technical Field

[0001] The present invention relates to the technical field of research on Bursaphelenchus xylophilus, in particular to an integrated sky-ground monitoring method for monitoring Bursaphelenchus xylophilus disease. Background Art

[0002] Remote sensing monitoring of dead trees caused by Bursaphelenchus xylophilus disease mainly relies on the change in the color of the tree crown, which has typical characteristics in visible light and multi-spectral remote sensing images. For the monitoring of discolored pine trees based on satellite remote sensing, the current satellite remote sensing monitoring methods for discolored pine trees mainly include pixel-based image classification methods, object-oriented methods, deep learning methods, vegetation index methods, etc. Currently, for monitoring the epidemic situation of Bursaphelenchus xylophilus disease using medium and low-resolution images, due to the discreteness of the spatial distribution of infected trees, the monitoring effect is very poor. For high-resolution images, domestically, mainly the domestic sub-meter GF-2 and BJ-2 satellite data are used to carry out the monitoring of Bursaphelenchus xylophilus disease. Research and practice have shown that by using methods such as supervised classification and object-oriented CART decision tree classification, combined with terrain data, the epidemic areas or diseased woods of Bursaphelenchus xylophilus disease have been extracted, and certain effects have been achieved. This kind of remote sensing data can meet the identification of typical red symptoms of coniferous leaves and large trees (or tree clusters) with a crown diameter of more than 5m, but it is still difficult to identify individual diseased woods; it is suitable for quickly and macroscopically monitoring the spatial distribution of the epidemic situation, and can provide a scientific basis for subsequent detailed investigations and the determination of the epidemic situation.

[0003] In addition, the unmanned aerial vehicle (UAV) remote sensing technology developed in recent years has brought solutions for the rapid positioning and total quantity estimation of discolored pine trees in China. However, currently, the extraction of UAV images of scattered individual discolored pine trees still remains at the level of visual interpretation. The working scheme that completely relies on manual visual interpretation of discolored pine trees has low efficiency and strong subjectivity. Although existing remote sensing data classification and extraction algorithms such as support vector machines, object-oriented methods, and neural networks have improved the efficiency of manual visual interpretation to a certain extent, in terms of time complexity, they are difficult to meet the calculation requirements of ultra-high spatial resolution large-area UAV images in the order of GB or even TB.

[0004] How to distinguish discolored pine trees from other red broad-leaved trees, sparse vegetation, bare soil, etc. is still a difficult problem faced by the current analysis based on visible light and multi-spectral data. The emergence of hyperspectral has provided a solution to the situation of same object with different spectra or different objects with the same spectrum on high-resolution images, but it still faces the dilemmas of large high-dimensional data volume and strong correlation between bands. Hyperspectral remote sensing data has continuous ground object spectral information. Utilizing this advantage, airborne hyperspectral remote sensing is used to identify the damage degree of Bursaphelenchus xylophilus disease, and the results are significantly better than multi-spectral images. However, due to the very high spatial resolution of airborne hyperspectral and the differences in local field of view conditions, lighting conditions, and canopy structures during data collection, the canopy brightness values of the same tree vary significantly, which brings some interference to the determination of discolored pine tree pixels. Moreover, the acquisition cost of hyperspectral remote sensing images is high and the data processing difficulty is large. Currently, it is only limited to the small-area research stage and has not been widely promoted on a large scale.

[0005] In summary, with the diversification of remote sensing platforms, the continuous improvement of remote sensing data quality, and the continuous progress of classification technologies, the theory and technology of pine wilt disease monitoring have developed rapidly. From pixel-based classification to object-oriented classification methods, from traditional classification methods to deep learning recognition methods, and from satellite remote sensing data sources to multi-platform and multi-spatial resolution data sources, it provides a basis for accurate classification and recognition. However, these methods are still at the level of single technology utilization in practice. The advantages of satellite remote sensing and UAV remote sensing have not been fully utilized, and a complete application technology system for production has not been formed. Cost and efficiency are still the main problems. Summary of the Invention

[0006] Aiming at the deficiencies of the existing technology, the present invention provides an integrated sky-ground monitoring method for pine wilt disease monitoring, which has high efficiency in screening epidemic areas, can achieve large-area and high-efficiency monitoring, is convenient for obtaining UAV remote sensing data, has high recognition accuracy, and is guaranteed for accuracy verification.

[0007] To achieve the above object, the present invention is realized through the following technical solutions: An integrated sky-ground monitoring method for pine wilt disease monitoring, comprising the following steps:

[0008] S1. Select the regional location. First, extract the infected dead trees through UAV images, extract the patches of infected dead trees through satellite images, and conduct on-site marking, positioning, and measurement of the dead trees through ground surveys.

[0009] S2. Obtain high-spatial-resolution satellite remote sensing data and UAV remote sensing data as data sources respectively, and combine the ground survey data with the second-class forest resource survey data to jointly form a data set of the monitoring area.

[0010] S3. Establish samples of diseased wood through the above data. The samples include the spectral characteristics, texture characteristics, and geometric characteristics of pine wilt disease-infected dead trees.

[0011] S4. According to the samples, the satellite remote sensing confirms the patches of infected dead trees, comprehensively uses image enhancement and image classification methods, and conducts verification to establish the spatial distribution information and location information of pine wilt disease-infected dead trees.

[0012] Preferably: The high-spatial-resolution satellite remote sensing data includes panchromatic band (470 - 830 nm) data with a spatial resolution of 50 cm and multi-spectral data of four bands of blue, green, red, and near-infrared with a spatial resolution of 2 meters.

[0013] Preferably: The UAV remote sensing data includes selecting one 1 km 2 - 2 km 2The selected area is used as the verification area, and the selected areas are photographed by drones respectively.

[0014] Preferably, the ground survey data includes field handheld GPS dead tree marking, positioning and measurement of dead trees in one or more partial sub-compartments selected within the area after drone photography to complete modeling and thematic analysis.

[0015] The data measured on-site includes: sub-compartment number, diseased wood number, diameter at breast height, crown width, visible crown width in the air, pine needle retention rate, crown color, coordinate x (GPS), coordinate y (GPS), coordinate x (image) and coordinate y (image). Through on-site investigation, an image feature library of diseased wood is established, including color, shape, texture, pattern, and spatial distribution characteristics.

[0016] Preferably, the forest resource class-II survey data includes spatial information of administrative boundaries and sub-compartment boundaries in each area, names of administrative regions at all levels, sub-compartment numbers, sub-compartment areas, land types, and tree species attribute information.

[0017] Preferably, the analysis of drone remote sensing data is used to extract the number of diseased dead trees. According to the visual characteristics of diseased wood in the images, the pine trees infected with pine wilt disease are selected for recording to determine the distribution results of diseased wood. The spatial resolution of drone images is very high. In the early stage of the disease, the diseased wood appears yellow in the true color images of the drone, and shows obvious reddish-brown or brown-yellow in the middle and late stages, which is significantly different from the green of healthy trees. For pine trees that have been diseased for half a year or longer, the needles fall off, and the branches appear white or grayish-white in the images. Based on the two characteristics of tree shape pattern and color, after manual interpretation of the diseased wood in the entire study area, preliminary results are formed. For some diseased wood with doubtful interpretation, further field investigation is required to proofread and verify the results, and finally the accurate distribution results of diseased wood on the drone images of the study area are obtained.

[0018] Preferably, the analysis of high-spatial-resolution satellite remote sensing data is used to extract patches of diseased dead trees. Based on satellite remote sensing images, first, the HSV threshold method is used to identify diseased wood patches, then the forest resource class-II survey background data is used to extract the distribution area of pine forests, and the forest gaps are removed through a segmentation algorithm. A diseased wood number model is established using the patch area and the number of diseased wood plants, and finally, the number of diseased wood plants in all patches is estimated using the number model, and the position coordinates of each plant are extracted; the segmentation algorithm starts from a single pixel and gradually merges upward into larger objects until the set segmentation scale (f) is met.

[0019] The segmentation scale (f) consists of four parameters, namely spectral heterogeneity (h color ), shape heterogeneity (h shape ), spectral information weight (w color ), and shape information weight (w shape), the sum of the weights of the spectral feature and the shape feature is 1 (i.e., w color + w shape = 1), f = w × h color + (1 - w) × h shape ;

[0020] The spectral heterogeneity (h color ) is not only related to the number of pixels of the composed object, but also depends on the standard deviation of each band: is the standard deviation of the pixel values inside the object, calculated according to the pixel values of the composed object, and n is the number of pixels;

[0021]

[0022] In addition, the shape heterogeneity (h shape ) is calculated by the compactness (h compact ) and the smoothness (h smooth ). The smoothness is used to optimize the smoothness of the boundary of the segmented object and can suppress the fragmentation of the edge; the compactness is used to optimize the compactness of the segmented object. The sum of the weights of the two indexes is also 1 (i.e., w compact + w smooth = 1);

[0023] h shape = w compact × h compact + (1 - w compact ) × h smooth ;

[0024] On the basis of image preprocessing, the P satellite image is segmented with the help of eCognition Developer software. The image is segmented at multiple scales, and the segmentation range is 20 - 150. The quantitative evaluation of the multi-scale segmentation results is carried out to find the optimal segmentation scale, and the optimal segmentation scale is selected through visual evaluation;

[0025] After quantitatively evaluating the optimal segmentation scale, the characteristic variables of each object in the object layer at the optimal segmentation scale in the P satellite image are exported through eCognition Developer, including spectral, texture, geometric characteristic variables and various indexes calculated from the original image bands;

[0026] For the RGB in the high-spatial-resolution satellite remote sensing data, where RGB represents the colors of the red, green, and blue three channels, the HSV transformation is performed on the multi-spectral data of the red, green, and blue 3 bands for image color enhancement, and the HSV color model is converted. There are obvious differences in the values of the diseased wood and healthy trees in the H band. By finding the threshold of the discriminator, the patches of the diseased wood can be automatically identified;

[0027] The RGB to HSV conversion formula is as follows:

[0028] V = max(R, G, B)

[0029]

[0030] If H < 0 then H = H + 360. On output 0 ≤ V ≤ 1, 0 ≤ S ≤ 1, 0 ≤ H ≤ 360.

[0031] After the patches of diseased trees are identified, combined with the data of the second-class forest resources survey, the range of pine forests is cropped, the non-pine forest areas are excluded, and the patches of diseased trees in the pine forest areas are obtained;

[0032] Through the local geometric correction method, when correcting, taking the satellite data as the reference, the results extracted by the UAV are geometrically corrected, so as to use the results of UAV data extraction as the reference data to match and overlay with the satellite data. Based on the UAV images, satellite images and their respective interpretation and identification results, through the size, color, spatial distribution characteristics, patterns of diseased trees at the same position of the two and the positional relationship with the surrounding ground objects, comprehensively using the methods of comparative analysis and logical inference, accurately judge the position of the diseased trees, obtain the geometrically precise correction results, and realize the one-to-one correspondence between the positions of the diseased trees interpreted from the UAV images and the identification results of the satellite images;

[0033] Through the interpretation results of UAV images and satellite patch data, a new vector data of satellite diseased tree points is obtained in a human-computer interaction manner, and this data is used for the modeling of the number of identified diseased trees;

[0034] Based on the satellite images, patches of diseased trees of different sizes are identified, and the number of diseased trees is calculated, which is obtained through the relationship model between area and number of trees. The model form is:

[0035] y = ax + b

[0036] Or

[0037] y = ax 2 + bx + c

[0038] Among them, y is the number of diseased trees, x is the patch area, and the number of diseased trees of all patches is solved by this model formula.

[0039] Preferably: Select some areas in the study area, conduct UAV shooting in the whole area or local sampling areas within one week after the satellite images are obtained, and obtain the positions of dead trees through the manual interpretation of UAV images, so as to verify the accuracy of the satellite image identification results;

[0040] The UAV images randomly distributed in multiple verification areas, through the manual interpretation results, obtain the number of trees in each area and the position of each dead tree in each area, which are respectively used for the verification and evaluation of position accuracy and number of trees accuracy.

[0041] Preferably: The plant number accuracy is verified based on UAV images, and the plant number accuracy p n is expressed as a function of the plant number error rate:

[0042] p n = 1 - |E n |

[0043]

[0044] where E n is the plant number error rate, n is the recognized plant number, m is the verified plant number, subscript i is the area number, n i is the recognized plant number in the i-th area, m i is the verified plant number in the i-th area, and b is the number of areas. E n can be positive or negative. When E n > 0, it means the recognized value is greater than the verified value; when E n < 0, it means the recognized value is less than the verified value; the verification information of the number of dead trees is obtained after manual ground inspection and control work to verify the accuracy of the recognition result;

[0045] The single-plant position accuracy error is represented by the Euclidean distance between the recognized point position (xd, yd) and the verified point position (xt, yt) For the verification area, the overall position error (Ep) is represented by the arithmetic mean of the single-plant position errors:

[0046]

[0047] where represents the x coordinate value of the i-th plant of the recognition data, represents the x coordinate value of the i-th plant of the verification data, represents the y coordinate value of the i-th plant of the recognition data, represents the y coordinate value of the i-th plant of the verification data; with the mean value E p as the main reference, and the maximum and minimum values are also considered.

[0048] Preferably: Using the image classification method, the open spaces (including forest edges) in the pine forest sub-compartments are separated from the forest to exclude the areas that are easily confused with diseased dead trees from the satellite image recognition results.

[0049] The beneficial effects of the present invention are:

[0050] It has high efficiency in screening epidemic areas, can achieve large-area and high-efficiency monitoring, uses remote sensing technology, and the recognition accuracy remains above 80% (error below 20%). Secondly, it realizes the identification of dead trees based on satellite remote sensing and extracts the position coordinates of each plant, obtaining information quickly and in detail, taking into account both efficiency and cost. Description of the Drawings

[0051] Figure 1 Schematic diagram of the accuracy verification sampling area in the present invention;

[0052] Figure 2 Flow chart of the multi-scale segmentation algorithm in the present invention;

[0053] Figure 3 Schematic diagram of the patch recognition result on the satellite remote sensing image in the present invention;

[0054] Figure 4 Distribution map of pine forests and forest clearings in Jinbei Sub-district in the present invention;

[0055] Figure 5 Distribution map of patches of diseased trees extracted by satellite remote sensing in Jinbei Sub-district in the present invention;

[0056] Figure 6 Local geometric precision correction map of the location of diseased trees interpreted from UAV images in the present invention;

[0057] Figure 7 Patch map of UAV images in the present invention;

[0058] Figure 8 Recognition and interpretation results of UAV images in the present invention;

[0059] Figure 9 Distribution map of diseased trees verified by local sub-compartments in the present invention. Detailed implementation manners

[0060] Example 1

[0061] As Figures 1-9 The integrated sky-ground monitoring method for pine wilt disease monitoring described above, comprising the following steps:

[0062] S1. Select the regional location. First, extract the diseased dead trees through UAV images, extract the patches of diseased dead trees through satellite images, and conduct on-site marking, positioning and measurement of the dead trees through ground surveys;

[0063] S2. Obtain high-spatial-resolution satellite remote sensing data and UAV remote sensing data as data sources respectively, and combine the ground survey data and the forest resource second-class survey data to jointly form a data set of the monitoring area;

[0064] The high-spatial-resolution satellite remote sensing data includes panchromatic band (470 - 830nm) data with a spatial resolution of 50cm and multi-spectral data of four bands of blue, green, red, and near-infrared with a spatial resolution of 2m.

[0065] The relevant technical parameters of the high-spatial-resolution satellite remote sensing data are shown in the following table;

[0066]

[0067] The UAV remote sensing data includes selecting an area of 1 km 2 - 2 km 2 in each of multiple areas as the verification area. Refer to the appendix Figure 1 In this embodiment, Shangdong Village, Jinma Village, and Xishu Neighborhood Committee with more pine forests and more severe epidemics are selected in Jinbei Sub-district, Lin'an District as verification villages. An area of about 1 km 2 - 2 km 2 is selected in each village as the verification area, and the selected areas are photographed by UAV respectively.

[0068] The technical parameters and area of the UAV images in the verification area are shown in the following table:

[0069]

[0070] Visible light images are taken by UAV in this area, and the orthophoto image spatial resolution is 3 cm - 7 cm.

[0071] The ground survey data includes field handheld GPS dead tree marking, positioning, and measurement of dead trees in some small compartments in one or more selected areas after UAV shooting to complete modeling and thematic analysis;

[0072] The data measured on the ground includes: small compartment number, diseased wood number, diameter at breast height, crown width, visible crown width in the air, retention rate of pine needles, crown color, coordinate x (GPS), coordinate y (GPS), coordinate x (image), and coordinate y (image). Through field investigation, an image feature library of diseased wood is established, including color, shape, texture, pattern, and spatial distribution characteristics.

[0073] The forest resource class II survey data includes the spatial information of administrative boundaries and small compartment boundaries in each area, the names of administrative regions at all levels, small compartment numbers, small compartment areas, land types, and tree species attribute information.

[0074] S3. Establish samples of diseased wood through the above data. The samples include the spectral characteristics, texture characteristics, and geometric characteristics of dead trees infected with pine wood nematode disease;

[0075] S4. According to the samples, satellite remote sensing is used to confirm the patches of infected dead trees. Image enhancement and image classification methods are comprehensively used and verified to establish the spatial distribution information and location information of dead trees infected with pine wood nematode disease.

[0076] The analysis of UAV remote sensing data is used to extract the number of diseased and dead trees. According to the visual characteristics of the diseased wood in the image, the pine trees infected with pine wilt disease are selected for recording to determine the distribution results of the diseased wood. The spatial resolution of UAV images is very high. In the early stage of the disease, the diseased wood appears yellow in the true color image of the UAV, and shows obvious reddish-brown or yellowish-brown in the middle and late stages, which is significantly different from the green of healthy trees. For pine trees that have been diseased for half a year or longer, the needles fall off, and the branches appear white or grayish-white in the image. Based on these two characteristics of the tree shape pattern and color, after the human-computer interaction interpretation of the diseased wood in the entire study area, preliminary results are formed. For some diseased wood with doubtful interpretations, it is necessary to conduct further field investigations to proofread and verify the results, and finally obtain the accurate distribution results of the diseased wood on the UAV images of the study area.

[0077] In this embodiment, the source of UAV remote sensing data is selected from Jinbei Sub-district, Lin'an District, Hangzhou City. The administrative area is 81.54 Km2 (122,310 mu), and the main tree species in Jinbei Sub-district is pine species dominated by Masson pine. There are 689 small forest plots with more than 10% pine forests in the area, with an area of 2,343.3 hm2 (35,150 mu), and the pine forest area accounts for 28.28% of the administrative area.

[0078] Reference appendix Figure 1 , in order to effectively verify the results of satellite remote sensing image machine recognition, one verification sampling area (referred to as the verification area) is set up in Shangdong Village, Jinma Village, Longma Village and Xishu Neighborhood Committee respectively. UAV images are taken of this area within one week after the satellite image shooting period, which is used as the basic data for verifying the recognition accuracy based on satellite images.

[0079] The analysis of high spatial resolution satellite remote sensing data is used for the extraction of diseased and dead tree patches. Based on satellite remote sensing images, first, the HSV threshold method is used to identify the diseased wood patches, then the forest resource second-class survey background data is used to extract the pine forest distribution area, and the forest gaps are removed through the segmentation algorithm. An epidemic tree number model is established using the patch area and the number of epidemic trees, and finally, the number of epidemic trees in all patches is estimated using the tree number model, and the position coordinates of each tree are extracted; the segmentation algorithm starts from a single pixel and gradually merges upward into larger objects until the set segmentation scale (f) is met;

[0080] The segmentation scale (f) consists of four parameters, namely spectral heterogeneity (h color ), shape heterogeneity (h shape ), spectral information weight (w color ) and shape information weight (w shape ). The sum of the weights of spectral features and shape features is 1 (i.e., w color +w shape = 1), and f = w×h color +(1 - w)×h shape; Spectral heterogeneity (h color ) is not only related to the number of pixels that make up the object, but also depends on the standard deviation of each band: is the standard deviation of the pixel values within the object, calculated based on the pixel values that make up the object, and n is the number of pixels;

[0081]

[0082] In addition, shape heterogeneity (h shape ) is calculated from compactness (h compact ) and smoothness (h smooth ). Smoothness is used to optimize the smoothness of the boundary of the segmented object and can suppress edge fragmentation; compactness is used to optimize the compactness of the segmented object. The sum of the weights of the two indicators is also 1 (i.e., w compact +w smooth = 1);

[0083] h shape = w compact ×h compact +(1 - w compact )×h smooth ;

[0084] On the basis of image preprocessing, the P satellite image is segmented with the help of eCognition Developer software. The image is segmented at multiple scales, and the segmentation range is 20 - 150. The P satellite image is segmented at multiple scales; as can be seen from the above algorithm, in the same area, as the segmentation scale increases, the number of segmented objects decreases, and the number of objects directly affects the operation speed and classification accuracy. When the segmentation scale is too low, the number of objects increases significantly, and the operation speed will be greatly slowed down. On the contrary, when the segmentation scale is too high, the number of objects decreases, and it is easy to cause different ground features to be segmented into one object, thus reducing the classification accuracy. Therefore, it is particularly important to quantitatively evaluate the multi-scale segmentation results and find the optimal segmentation scale. Through visual evaluation, the optimal segmentation scale is selected as 100.

[0085] After quantitatively evaluating the optimal segmentation scale, the characteristic variables of each object in the object layer at the optimal segmentation scale in the P satellite image are exported through eCognition Developer, including spectral, texture, geometric characteristic variables, and various indices calculated from the original image bands;

[0086] Currently, the methods for texture extraction mainly include four methods: based on statistical description, based on wavelet transform, based on application of fractal theory, and based on geostatistics. Among them, the gray-level co-occurrence matrix (GLCM) based on statistical description has been proven to play an important role in vegetation classification, especially the homogenization of the gray-level co-occurrence matrix (GLCM_HOMO). Therefore, in this embodiment, the gray-level co-occurrence matrix algorithm is selected to extract the texture information of the object.

[0087] Spectral features are the most important features for remote sensing image classification. There are significant differences in the spectral features of healthy trees and diseased trees in the blue, green, red, and near-infrared bands. Therefore, in this embodiment, the spectral feature variables calculated for the object include the mean and standard deviation of each band, and the normalized difference vegetation index (NDVI).

[0088] Geometric attributes mainly describe the shape and size of the object. When used to analyze wetland features in areas with a relatively small scale or less human activity interference, the overall effect is lower than that of spectral and texture features. This can be explained to a certain extent by the small co-dependence between objects.

[0089] Specific variables such as spectral, texture, and geometric features are shown in the following table:

[0090]

[0091] For the RGB in high-spatial-resolution satellite remote sensing data, where RGB represents the colors of the red, green, and blue channels,

[0092] various colors are obtained by changing the red (R), green (G), and blue (B) color channels and their superposition with each other. This standard includes almost all colors that can be perceived by human vision and is one of the most widely used color systems at present.

[0093] HSV: HSV (Hue, Saturation, Value) is a color space created by A.R. Smith in 1978 based on the intuitive characteristics of colors, also known as the Hexcone Model. The HSV color model can be transformed from the RGB model. The parameters of colors in this model are: hue (H), saturation (S), and value (V). The value range of H is from 0 degrees to 360 degrees, where 0 degrees represents red, and the higher the degree, the closer it is to green. S is expressed as a percentage to represent saturation. The larger the percentage, the more saturated the color is. That is, when the S value is large, the saturation is high, and the image color is also darker; when the S value is small, the saturation is low, and the image color is lighter. V is also expressed as a percentage, 0% represents black, and 100% represents white.

[0094] Refer to Appendix Figure 3 , the color characteristics of diseased trees presented in the satellite remote sensing image with a resolution of 0.5 meters are similar to those of drones, but the colors are darker. By performing HSV transformation on the multi-spectral data of the three red, green, and blue bands for color enhancement and converting to the HSV color model, there are obvious differences in the values of the H band between diseased trees and healthy trees. By finding the threshold of the discriminator, the patches of diseased trees can be automatically identified.

[0095] The RGB to HSV conversion formula is as follows:

[0096] V=max(R,G,B)

[0097]

[0098] If H<0then H=H+360.On output 0≤V≤1,0≤S≤1,0≤H≤360.

[0099] After the diseased wood patches were identified, the pine forest range was trimmed and non-pine forest areas were removed in combination with the second-category forest resource survey data to obtain diseased wood patches in the pine forest area.

[0100] When the diseased tree plaques extracted by satellite images are superimposed and analyzed with the results of drone image interpretation, the spatial reference basis of the diseased tree positions of the two is required to be consistent, and neither of them can have any deformation. In this embodiment, the diseased tree point vector results of the drone image of the same slope with relatively consistent deviation direction are matched with the satellite data through a local geometric correction method. During the correction, the satellite data is used as a reference to perform geometric correction on the results extracted by the drone, so that the results of drone data extraction can be used as reference data for matching and superimposed analysis with the satellite data. Based on the drone image, satellite image and their respective interpretation and recognition results, the diseased tree size, color, spatial distribution characteristics, pattern and positional relationship with surrounding objects at the same position of the diseased tree are used to comprehensively use comparative analysis and logical inference methods to make an accurate judgment on the diseased tree position, obtain a geometric precision correction result, and achieve a one-to-one correspondence between the position of the diseased tree interpreted by the inorganic image and the satellite image recognition result;

[0101] By superimposing the diseased tree point data interpreted by drone images with the diseased tree point data identified by satellite images, it is found that after geometric correction, most of the diseased tree point data interpreted by drone images fall into the diseased tree patches identified by satellite images, but a small number of diseased tree patches have no corresponding points in the drone data. After observing and comparing the drone image data with the satellite image data, it is found that this part of the map contains deciduous diseased trees, healthy broad-leaved trees with red leaves, open spaces and wastelands in the forest, and some diseased trees under the forest with a large degree of tree crown coverage. By adding attribute descriptions to the diseased tree patches extracted from satellite images, the diseased trees under the forest that were missed during drone image interpretation are supplemented, and new drone diseased tree interpretation point vector data are generated for verification.

[0102] Most of the patches of diseased trees extracted from satellite images contain only one diseased tree per patch, but there are also many patches with two or more diseased trees, and even some contiguous patches with 7 - 8 diseased trees. Among the contiguous patches with two or more diseased trees, there are some diseased trees with partial fallen leaves. To verify and analyze the accuracy of the identification of UAV images, for the patches in the above two situations, combined with the interpretation results of UAV images, a supplementary interpretation of the diseased tree situation corresponding to the patches was carried out, and the attributes were recorded. Finally, all the result patches automatically identified by satellite data were recorded with attributes, and the attributes included five situations, namely diseased trees, understory diseased trees, fallen-leaf diseased trees, broad-leaved trees, and empty land with weeds. The patch data was used to evaluate the position accuracy of diseased tree identification. At the same time, referring to the interpretation results of UAV images and satellite patch data, a new satellite diseased tree point vector data was obtained through a human-computer interaction method. This data was used for the modeling of the number of diseased trees identified.

[0103] Based on the satellite images, diseased tree patches of different sizes were identified, and the number of diseased trees was calculated through a relationship model between area and number of trees. The model form is:

[0104] y = ax + b

[0105] Or

[0106] y = ax 2 + bx + c

[0107] Among them, y is the number of diseased trees, and x is the patch area. The number of diseased trees in all patches was solved using this model formula.

[0108] In this embodiment, the model was established on a village-by-village basis. The modeling samples (patches) were randomly selected from each village as a whole. The number of sample units selected from each population was not less than 2% of the total population (the number of patches), and the minimum number of sample units was not less than 200.

[0109] For all sample units (the selected patches), the number of diseased trees in each sample unit was visually interpreted based on UAV images. Using the interpretation results as the true values, a linear or polynomial relationship model between the interpreted number of trees in each patch and its area was established. When the determination coefficient r2 of the model is greater than 0.60, the model is valid, and the number of diseased trees in all patches was solved using this model.

[0110] In a part of the study area, within one week after the satellite image acquisition date, UAV photography was carried out in the whole area or a partial sampling area. The positions of dead trees were obtained through the manual interpretation of UAV images to verify the accuracy of the satellite image identification results;

[0111] The UAV images randomly distributed in multiple verification areas. In this embodiment, the UAV images in multiple verification areas (A, B, C, D) randomly distributed in Jinbei Street are used. The number of plants in each area and the location of each dead tree in each area are obtained through manual interpretation results, which are used for the verification and evaluation of location accuracy and plant number accuracy respectively.

[0112] The plant number accuracy is verified based on the UAV images. The plant number accuracy p n is expressed by the function of the plant number error rate:

[0113] p n = 1 - |E n |

[0114]

[0115] where E n is the plant number error rate, n is the recognized plant number, m is the verified plant number, the subscript i is the area number, n i is the recognized plant number in the i-th area, m i is the verified plant number in the i-th area, and b is the number of areas. E n can be positive or negative. When E n > 0, it means the recognized value is greater than the verified value; when E n < 0, it means the recognized value is less than the verified value; the verification information of the dead tree quantity is obtained after the manual ground inspection and control work to verify the accuracy of the recognition result;

[0116] The single-plant position accuracy error is represented by the Euclidean distance between the recognized point position (xd, yd) and the verified point position (xt, yt) For the verification area, the overall position error (Ep) is represented by the arithmetic mean of the single-plant position errors:

[0117]

[0118] where represents the x coordinate value of the i-th plant in the recognized data, represents the x coordinate value of the i-th plant in the verified data, represents the y coordinate value of the i-th plant in the recognized data, represents the y coordinate value of the i-th plant in the verified data; taking the mean value E p as the main reference, and considering the maximum and minimum values at the same time.

[0119] The image classification method is adopted to separate the open spaces (including forest edges) in the small pine forest classes from the forest, so as to eliminate the areas that are easily confused with diseased dead trees from the satellite image recognition results.

[0120] Refer to Appendix Figure 4, Through the above method, the results of diseased tree classification show that there are 2,146 forest clearings with a total area of 185.04 hm² (2,776 mu), and the average area of each clearing is 862 m 2 .

[0121] Refer to the appendix Figure 5 . In this embodiment, based on satellite remote sensing images, a distribution map of diseased tree patches is extracted. After H-threshold processing, the patch distribution of the spatial distribution of diseased trees is raster data. To facilitate the extraction of the coordinate position information and the number of trees of diseased trees, the results are vectorized. After removing the forest clearings, the vector data is checked against the UAV images, and it is found that there are basically no diseased trees in patches smaller than 1 square meter, and the patches with an area smaller than 1 square meter are removed. Calculate the area and the center point of each patch. There are 11,310 diseased tree patches in Jinbei Sub-district, and the total area of the diseased tree patches is 180,860 m² (217.3 mu).

[0122] Refer to the appendix Figures 5-6 . In this embodiment, due to the different correction accuracies of UAV images and satellite images and the large change in terrain height difference of UAV images, before building a model for the number of diseased trees, it is necessary to accurately register the recognition results (dead tree patches) based on satellite images with the interpretation results (single dead trees) based on UAV images. Select an area with consistent small terrain in a 56 hm 2 pine forest in the northern part of Shangdong Village, and use the spline function model for sub-area fine correction. By selecting 10 sub-compartments in the northern part of Shangdong Village and using the local geometric correction method, referring to the UAV images and satellite images, geometric fine correction is performed on the positions of dead trees interpreted based on UAV images. Finally, among the 1,385 plants that exactly correspond to the satellite image recognition patches in the interpretation results of UAV images, the number of patches is 935.

[0123] The relationship diagram between diseased tree patches and the corresponding number of trees is shown in the following table;

[0124] ID Small class number <![CDATA[Patch area (m 2 )]]> Number of trees 1 5 14.0 2 2 5 3.3 1 3 5 12.0 1 4 1 3.1 1 5 5 8.4 1 6 5 60.8 3 7 5 16.8 1 8 1 4.3 1 …… …… …… …… 934 47 45.0 2 935 47 20.0 1

[0125] Using the interpretation results (plants) of UAV images after geometric fine correction, and comparing with the satellite image recognition results (patches), confirm the number of diseased trees (dead trees) in each patch. Finally, obtain a patch information confirmation form with the patch number, patch area, and the number of diseased trees as the main contents. Based on this form, establish a relationship model between the area and the number of trees.

[0126] The following figure is the linear model of Shangdong Village. The relationship between the area of diseased tree patches and the number of trees can be obtained: y = 0.0206x + 0.785 (y - the number of trees, x - the area), and the coefficient of determination is 0.6422. Based on this model, use the patch data automatically recognized and extracted by satellites to estimate the number of diseased trees in each patch.

[0127]

[0128] According to the relationship model between the area and the number of trees of diseased wood patches, using the extracted diseased wood patch data, estimate the number of diseased wood trees in Jinbei Sub-district. According to the statistics of the second-class forest resource inventory plot data, the administrative area of Jinbei Sub-district is 8,019.3 hm² (120,289 mu), there are 689 pine forest plots, and the total area of pine forest plots is 35,150 mu.

[0129] According to the patch recognition results, through the extraction of epidemic areas, combined with interference analysis and the modeling and estimation of the number of diseased wood trees, it is found that there are 12,814 dead pine trees infected with Bursaphelenchus xylophilus in Jinbei Sub-district. Calculated according to the total administrative area, the average is 0.11 trees per mu; calculated according to the actual pine forest (epidemic-related) area, the average is 0.37 trees per mu. See the following table for details.

[0130]

[0131] In this embodiment, the accuracy evaluation is carried out by means of typical sampling, and the sample characteristic value is an effective estimate of the overall characteristic value it represents. The overall tree number accuracy (correct rate) and location accuracy of Jinbei Sub-district are represented by the mean of the tree number accuracy (correct rate) and location accuracy of the 4 sample areas selected from Shangdong Village, Jinma Village and Xishu Sub-district.

[0132] The evaluation of tree number accuracy is carried out by using the degree of difference in the number of trees between the machine recognition result and the visual interpretation result of the evaluator. According to the following formula:

[0133] p n =1 - |E n |

[0134]

[0135] Calculate the tree number error and tree number accuracy respectively.

[0136] In the evaluation of tree number accuracy, the recognition and interpretation data are obtained by means of typical sampling. Considering the representativeness and stability of the samples, all pine-containing sample units (plots) in four verification areas A, B, C, and D are selected. The actual area of pine forest plots in the verification areas is 217.7 hectares (3,266 mu), and the number of plots is 46. It is concluded that the recognition accuracy of the number of dead trees in Jinbei Sub-district is: 80.2%. See the following table for details.

[0137]

[0138] The evaluation of location accuracy is carried out by using the degree to which the machine recognition location deviates from the visual interpretation location of the evaluator. In this embodiment, in a typical (severely diseased) area, according to the smallest forest division unit, a plot with a medium area - Plot No. 15 of Shiqiao in Shangdong Village is selected as the sample for testing the location accuracy of the machine recognition result of satellite images, and the mean value of the location error is calculated by formula 3.8 for the overall location error analysis.

[0139] Reference appendix Figure 8 For inspection, the visual interpretation method was adopted. The inspector interpreted the dead trees in the area of Plot 15 on the satellite image and made corrections with reference to the UAV images taken during the same period. Finally, the corrected results were compared and analyzed with the machine recognition results of Plot 15 under inspection. The analysis was carried out using the neighborhood analysis tool of ArcMap.

[0140] There were 105 dead trees identified by satellite image in Plot 15, and the result confirmed by manual comprehensive interpretation was 107. According to the minimum distance, the dead tree point pairs corresponding to each other between machine recognition and manual interpretation were determined, and the distance statistical features of each point pair were counted to reflect the position error. The results were as follows: the average value (position error) was 3.16 meters, the minimum value was 0.79 meters, and the maximum value was 8.05 meters. See the following table for details.

[0141]

[0142] Using the monitoring results of 10 plot areas near Shijiawu in Shangdong Village, the interference factors affecting the identification accuracy of diseased wood were further analyzed, and the actual type composition of dead trees (diseased wood) in the identification results was further clarified. The total area of this region was 56.5 hm² (848 mu), and 1,466 diseased wood were identified. Using the results of the UAV as auxiliary data, the satellite monitoring results were further interpreted and type-subdivided. A total of 1,385 dead trees (diseased wood) of various types were confirmed, and 81 were misjudged due to interference factors such as broad-leaved trees and forest land gaps;

[0143] Based on the visual interpretation results of the inspector, the overall accuracy of the machine recognition results of the satellite image was 86.5%. The recognition accuracy of 8 out of 10 plots was greater than 94.0%, and 1 was less than 50%. The statistical reference of the proportion of the satellite image interpretation and recognition result types in each plot is as follows in the table.

[0144]

[0145] The confirmed number of trees in Plot 47 was significantly less than the recognized number. After on-site comparison, the reason was that broad-leaved trees accounted for the absolute majority in this plot, and there were many deciduous tree species. In the satellite image taken on October 23, some broad-leaved trees had turned yellow, resulting in some broad-leaved trees with yellowed leaves being counted as dead trees in the automatic recognition. This also indicates that other discolored trees, especially discolored broad-leaved trees, are the main interference sources for the remote sensing identification of pine wilt disease.

[0146] Seven plots were selected in Shangdong Village, and the satellite remote sensing monitoring results were verified through the man-machine interactive interpretation of UAV images. The results showed that the monitoring accuracy of the number of infected dead trees was greater than 74.4%, and the number monitoring accuracy of more than 70% of the plots was greater than 67%. The verification results of the local satellite monitoring plots in Shangdong Village are shown in the following table.

[0147]

[0148] In Jinma Village, 15 small classes were selected for human-computer interactive interpretation of UAV images, and the results were used to verify the results of satellite remote sensing monitoring. The result showed that the overall accuracy of the number of diseased and dead tree strains monitored was 81.9%. There were significant differences in the accuracy of each small class, ranging from 25.0% to 96.9%, but the number of small classes with an accuracy exceeding 60.0% still accounted for the majority. For the small classes with large errors, most of their areas were small. The verification results of the satellite local monitoring small classes in Jinma Village are shown in the following table.

[0149]

[0150]

[0151] This technology of the present invention solves the bottleneck problem of low efficiency in the monitoring of pine wilt disease faced by current remote sensing and geographic information system technologies. Using remote sensing technology, the recognition accuracy remains above 80% (error below 20%). Secondly, it realizes the identification of dead trees based on satellite remote sensing and the extraction of the position coordinates of each tree, obtaining information quickly and in detail, taking into account both efficiency and cost. Thirdly, this technology system makes full use of existing tool software in aspects such as geographic information systems and remote sensing, ensuring the process flow of the process and the consistency of the output results.

[0152] Example Two

[0153] In this Example Two, within the small classes of pine forests, by checking the forest cover situation from satellite remote sensing images, it can be found that there are some areas within the small class areas that are not covered by forests, such as forest gaps, wastelands, and forest edge bare lands. The spectral characteristics of these ground objects are very similar to those of diseased wood. In order to reduce the interference of these ground objects on the extraction of diseased wood, this Example Two will further eliminate these non-forest-covered areas. The images of the pine forest small class areas are classified using the object-oriented combined with random forest classification method to distinguish the forest-covered areas from types such as wastelands and forest edge bare lands, thereby eliminating the non-forest areas within the pine forest small classes and excluding interference factors for the accurate extraction of diseased wood; the random forest model is a new machine learning algorithm based on decision trees. First, M new training sets are randomly sampled with replacement from the original dataset, and the sampling quantity is about 2 / 3 of the original dataset. Then, M attributes are randomly sampled from the new training set to generate decision trees. Finally, by aggregating the prediction results of N decision trees, the class of the new sample is determined by voting, and the internal error can be estimated using the 1 / 3 data that is not drawn each time. Among many machine learning algorithms, the random forest has the following three characteristics and advantages: First, it has excellent classification performance and can process large data without feature selection and deletion; second, there is little manual intervention, usually no data preprocessing is required, and it can determine the features to be used according to the data itself; third, it has a fast operation speed and is easy to perform parallel processing.

[0154] Example Three

[0155] Select the experimental results completed by this method in the general survey of pine wilt disease in Jindong District of Jinhua City and Yongkang City of Jinhua City in 2020.

[0156] First, remote sensing images were obtained by aerial photography in two areas with an area of 658.19 Km2 in Jindong District of Jinhua City and 1049 Km2 in Yongkang City of Jinhua City. The image shooting time was October 31, 2020. The spatial resolution of the aerial image was 0.5 meters, including four bands: red, green, blue, and near-infrared. The census results are shown in the following table.

[0157]

[0158] In the second step, the recognition results in Jindong District were verified.

[0159] Four townships, namely Jiangdong, Lingxia, Yuandong, and Chisong, were selected as samples in Jindong District. The census data based on the integrated space-air-ground method was verified with the actual number of felled diseased trees during the control process in this area, and the verification results were used as the verification results for the entire Jindong District. The verification referred to the account records of the diseased tree control results in this area from January to April 2021. The specific verification results are shown in the following table.

[0160]

[0161]

[0162] Generally, the error rate of the number of trees was 4.1%, showing that the identified number of trees was slightly more than the number of cleared trees, but the sub-item differences were large, up to -29.5%. The identified number of trees was much lower than the actual number of cleared trees. The main reason was that the mountainous areas in the remote mountainous areas of this township were high and the forests were dense, and the crowns of the diseased trees were covered more. It also showed from another aspect that the township had increased the control efforts in the remote mountainous areas. For the entire Jindong District, its error mean and extreme values should be within the range.

[0163] In the third step, the recognition results in Yongkang City were verified.

[0164] Two township-level units, namely Chengdong Sub-district and Economic Development Zone, were selected as samples in Yongkang City, and two methods were used to verify their recognition results.

[0165] The first was to verify based on the actual number of cleared diseased trees. Using the account records of the diseased tree control results in this area from January to May 2021 as the standard for the number of dead trees, the error rate of the number of trees in the same area was calculated. The specific verification results are shown in the following table.

[0166]

[0167] In the table, there are certain differences between the census data and the cleaning data of Chengdong Sub-district and the Economic Development Zone. Moreover, the differences in the Economic Development Zone are relatively large. However, generally, they fluctuate up and down around the census data. As the overall scale expands, the error rate decreases.

[0168] Example 4

[0169] Two township-level units, Chengdong Sub-district and the Economic Development Zone, are selected as samples in Yongkang City and verified by the second method. Based on the local verification of UAV images, Gedangshan Village, Gedangxia Village, and Xiadalu Village in Chengdong Sub-district are selected. High-spatial-resolution UAV images are used to manually interpret dead trees one by one, and the error rate of the number of trees determined based on the recognition results of aerial images is determined. The specific verification results are shown in the following table.

[0170]

[0171] The above embodiments are only used to explain the inventive concept of the present invention, rather than limiting the protection scope of the rights of the present invention. Any non-substantive modification made to the present invention using this concept shall fall within the protection scope of the present invention.

Claims

1. An integrated sky-ground monitoring method for monitoring pine wilt disease, characterized in that It includes the following steps: S1. First, extract diseased trees through drone images, extract patches of diseased trees through satellite images, and conduct on-site marking, positioning, and measurement of diseased trees through ground surveys; S2. Obtain high-spatial-resolution satellite remote sensing data and drone remote sensing data as data sources respectively, and combine ground survey data with the second-class forest resources survey data to jointly form a data set for the monitoring area; S3. Establish samples of diseased trees based on the above data. The samples include spectral characteristics, texture characteristics, and geometric characteristics of diseased trees; S4. Confirm patches of diseased trees according to the samples and satellite remote sensing, comprehensively use image enhancement and image classification methods, and conduct verification to establish the spatial distribution information and location information of diseased trees; The analysis of high-spatial-resolution satellite remote sensing data is used for the extraction of patches of diseased trees. Based on satellite remote sensing images, first use the HSV threshold method to identify patches of diseased trees, then use the second-class forest resources survey background data to extract the distribution area of pine forests, and use the segmentation algorithm to eliminate forest gaps. Establish a diseased tree number model with the patch area and the number of diseased trees, and finally use the number model to estimate the number of diseased trees in all patches and extract the position coordinates of each tree; The segmentation algorithm adopts a bottom-up strategy, starting from a single pixel and gradually merging to form larger objects until the set segmentation scale is met. The segmentation scale is set to f; The segmentation scale consists of four parameters, namely spectral heterogeneity h color , shape heterogeneity h shape , spectral information weight w color and shape information weight w shape . The sum of the weights of spectral information and shape information is 1, that is, w color +w shape =1, f = w color ×h color +(1 - w color )×h shape ; Spectral heterogeneity h color It is not only related to the number of pixels of the constituent object, but also depends on the standard deviation of each band: is the standard deviation of the pixel values within the object, which is calculated based on the pixel values of the constituent object; In addition, the shape heterogeneity h shape is calculated from the compactness h compact and the smoothness h smooth The smoothness is used to optimize the smoothness of the boundary of the segmented object; the compactness is used to optimize the compactness of the segmented object. The sum of the weights of the two metrics is also 1, i.e., w compact + w smooth = 1; h shape = w compact × h compact + (1 - w compact ) × h smooth ; On the basis of image preprocessing, segment the P satellite image with the help of eCognition Developer software, perform multi-scale segmentation on the image, with the segmentation range being 20 - 150, conduct quantitative evaluation on the multi-scale segmentation results, find the optimal segmentation scale, and select the optimal segmentation scale through visual evaluation; After quantitatively evaluating the optimal segmentation scale, export the characteristic variables of each object in the object layer at the optimal segmentation scale in the P satellite image through eCognition Developer, including spectral, texture, geometric characteristic variables, and various indices calculated from the original image bands; Perform HSV transformation on the RGB in the high-spatial-resolution satellite remote sensing data. RGB represents the colors of the red, green, and blue channels. Perform HSV transformation on the multi-spectral data of the three red, green, and blue bands to enhance the image color, convert to the HSV color model. Diseased trees and healthy trees show obvious differentiation in the value of the H band. By analyzing and determining the classification threshold for distinguishing diseased trees and healthy trees, the patches of diseased trees can be automatically identified.

2. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: The RGB to HSV conversion formula is as follows: V = max{R, G, B} If H < 0, then H = H + 360 Output: 0 ≤ V ≤ 1, 0 ≤ S ≤ 1, 0 ≤ H ≤ 360 After identifying the patches of diseased trees, combine the second-class forest resources survey data to crop the pine forest range, eliminate non-pine forest areas, and obtain the patches of diseased trees in the pine forest area; Through the local geometric correction method, when correcting, the satellite data is used as a reference to geometrically correct the results extracted by the drone, so as to use the results of drone data extraction as reference data for matching and overlay analysis with satellite data. Based on the drone images, satellite images and their respective interpretation and recognition results, through the size, color, spatial distribution characteristics, patterns of diseased trees at the same position of the two and the positional relationship with surrounding ground objects, the methods of comparative analysis and logical inference are comprehensively used to accurately judge the positions of diseased trees, obtain the geometrically precise correction results, and achieve a one-to-one correspondence between the positions of diseased trees interpreted from drone images and the recognition results of satellite images; Through the drone image interpretation results and satellite patch data, a new vector data of satellite diseased tree points is obtained in a human-computer interaction manner, and this data is used for modeling the number of identified diseased trees; Based on the satellite images, diseased tree patches of different sizes are identified, and the number of diseased trees is calculated, which is obtained through the relationship model between area and number of trees. The model form is: y = ax + b Or y = ax 2 + bx + c Among them, y is the number of diseased trees, and x is the patch area. The number of diseased trees of all patches is solved using this model formula.

3. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, wherein: The high-spatial-resolution satellite remote sensing data includes panchromatic band data with a spatial resolution of 50 cm and multi-spectral data of four bands, namely blue, green, red, and near-infrared, with a spatial resolution of 2 m. Among them, the range of the panchromatic band is: 470 - 830 nm.

4. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, wherein: The UAV remote sensing data includes data from multiple regions, and a region of 1 km 2 - 2 km 2 is selected as the validation area in each region, and the selected areas are photographed by the UAV respectively.

5. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: The ground survey data includes field handheld GPS marking, positioning, and measurement of dead trees in some sub-compartments in one or more selected areas within the area after the drone is photographed, in order to complete modeling and thematic analysis; The data measured on the ground includes: sub-compartment number, diseased tree number, diameter at breast height, crown width, visible crown width in the air, retention rate of pine needles, crown color, GPS coordinates x and y, and coordinates x and y with respect to the image. Through field surveys, a diseased tree image feature library is established, including color, shape, texture, pattern, and spatial distribution characteristics.

6. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: The forest resource class-II survey data contains the spatial information of administrative boundaries and sub-compartment boundaries in each region, as well as the names of administrative regions at all levels, sub-compartment numbers, sub-compartment areas, land types, and tree species attribute information.

7. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, wherein: The analysis of drone remote sensing data is used to extract the number of diseased trees. According to the visual characteristics of diseased trees in the images, the pine trees infected with pine wilt disease are selected for recording to determine the distribution results of diseased trees.

8. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: Extract some areas and conduct drone photography in the whole region or local sampling areas within one week after the satellite image acquisition date. The positions of dead trees are obtained through the manual interpretation of drone images to verify the accuracy of the satellite image recognition results; The drone images randomly distributed in multiple verification areas, through the manual interpretation results, obtain the number of trees in each area and the position of each dead tree in each area, which are respectively used for the verification and evaluation of position accuracy and number of trees accuracy.

9. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: Verification of plant number accuracy based on UAV images, plant number accuracy p n Expressed as a function of the plant number error rate: p n = 1 - |E n | Among them, E n is the tree number error rate, n i is the recognized tree number in the i-th area, m i is the verified tree number in the i-th area, b is the number of areas. When E n > 0, it means that the recognized value is greater than the verified value; when E n < 0, it means that the recognized value is less than the verified value; the verification information of the number of dead trees is obtained after the manual ground survey and control work to verify the accuracy of the recognition result; The single-plant position accuracy error is represented by the Euclidean distance between the identified point position (x d , y d ) and the verification point position (x t , y t ). For the verification area, the overall position error E is represented by the arithmetic mean of the single-plant position errors: p is represented by the arithmetic mean of the single-plant position errors: Among them, represents the x - coordinate value of the i - th strain of recognition data, represents the x - coordinate value of the i - th strain of verification data, represents the y - coordinate value of the i - th strain of recognition data, represents the y - coordinate value of the i - th strain of verification data; with the overall position error E p as the main reference, while considering the maximum and minimum values.

10. The integrated sky-ground monitoring method for pine wilt disease monitoring according to claim 1, characterized in that: Using the image classification method, the forest clearings in the pine forest sub-compartments are separated from the forest to eliminate the areas confused with diseased trees from the satellite image recognition results. The area covered by the forest includes the forest edge.

Images